How much context do your MCP servers cost?
Every MCP server you enable sends its tool definitions to the model before you type anything. Tick the servers you plan to keep on to see the sum and which to turn off first. Figures are tools_list_tokens_model from scored.json: 107 of 187 listings measured; 80 show not measured and add nothing.
Each enabled MCP server adds its tool definitions to the model's context. In our measurements (107 of 187 listings), that ranges from 47 tokens (Calculator, lowest non-deprecated) to 241,447 tokens (Microsoft 365 MCP (Softeria)). Tick your servers below to see the total.
- Lowest measured (non-deprecated): Calculator — 47 tokens, 1 tool(s).
- Highest measured: Microsoft 365 MCP (Softeria) — 241,447 tokens, 202 tool(s).
- Full ranking (same numbers): lowest context cost MCP servers.
- Config tip: the same picker lives in Use the same MCP servers in Claude Code, Codex and Cursor.
Context budget picker
Method: tiktoken cl100k_base over compact JSON of each tool's name, description and inputSchema (field tools_list_tokens_model). It is an approximation: real session counts differ by tokenizer and client wrapping. Deprecated servers (7) are included in the picker and marked on the ranking page.
Loading…
Ticked servers stay in the total even when the filter hides them. Clear unticks everything and empties the filter.
- Needs JavaScript. The full ranked list is on the lowest context cost ranking.
Which servers to turn off first
Your ticked servers, biggest context cost first. Each row shows its share of your total and how many tokens you would save by turning off that server and every one above it. Servers we have not measured are listed last and are not counted.
| Server | Tokens | Share of your total | Running saving | Lower-cost option | Measured on |
|---|
Method details
Tokens are tools_list_tokens_model: tiktoken cl100k_base over the compact JSON of each tool's name, description and inputSchema from the tools/list we captured. Share = that server's tokens divided by the total of your measured ticked servers, rounded to 0.1%. The running saving adds each row to the rows above it. Lower-cost options list up to 3 servers with a lower measured cost that are not deprecated. Servers we publish a side-by-side comparison for come first. After those come other measured servers in the same category, grouped under Same category, not a direct swap. Servers with no measurement are never suggested. If none qualify, the row shows a dash. Lower context cost does not mean it does everything the other one does. Measured on is the capture date for that server (latest 2026-10-31). Servers with no measurement are never given a number.
How the tokens are counted
Source field: tools_list_tokens_model in scored.json, computed with tiktoken cl100k_base over the compact JSON of name + description + inputSchema for each tool in the captured tools/list. Extra tools/list fields (for example outputSchema, icons) are tracked separately as tools_list_tokens_extra and are usually not sent to the model; they are not added here. Page built 2026-10-31; capture dates on measured rows go through 2026-10-31. We do not estimate missing values.
Claude Code documents a tool search mode that can load tools on demand rather than all up front (Anthropic MCP docs). Not tested here: each client's live in-session token count.
What we have not tested
- Each client's real token count in a live session (the picker uses one tokenizer for all).
- Whether a given client forwards extra tools/list fields to the model.
- 80 listings with no
tools_list_tokens_model— shown as not measured, not estimated.
FAQ
How many tokens does an MCP server add to context?
It depends on the server. Of the 107 listings we measured, the lowest non-deprecated is Calculator at 47 tokens and the highest is Microsoft 365 MCP (Softeria) at 241,447 tokens. Add up the servers you enable with the picker on this page.
How is context cost measured?
tools_list_tokens_model in scored.json: tiktoken cl100k_base over the compact JSON of each tool's name, description and inputSchema from the tools/list we captured. 107 of 187 listings have a measured value; the rest show not measured. Counts on this page were read from scored.json on 2026-10-31.
Why do some servers say not measured?
80 listings have no tools_list_tokens_model (no tools/list capture with at least one tool, or the handshake needed credentials). We do not estimate.
Is this the same number my client uses?
No. It is an approximation. Claude, GPT and other models use their own tokenizers, and each client adds wrapping. Some clients also load tools on demand rather than all up front (Claude Code documents a tool search mode).
Where is the full ranking?
The same measured numbers power our lowest context cost MCP servers ranking. The picker also appears inside the cross-client MCP config guide.
No product here is rated or ranked by quality, and this page has no affiliate links. How we test: methodology. Related: lowest context cost ranking · cross-client config guide.
Data read 2026-10-31 from scored.json